COSMOS is a city-scale advanced wireless testbed — real radios, real fibre, real rooftops in West Harlem, plus the indoor ORBIT grid and its sandboxes in New Jersey. These tutorials are the hands-on route into it: 71 guides across 14 topics, each one an experiment you can actually run rather than a description of one.
Three things apply to everything on this page, and are worth doing once:
| Step | Do this | Why |
|---|---|---|
| 1 | ORBIT Basics: the six-step experiment life cycle | The shape every experiment takes — reserve, image, configure, run, collect, release |
| 2 | Loading a disk image with omf load |
Nodes boot the image you give them. This is the single most-used command on the testbed |
| 3 | Getting started with COSMOS SDR resources | What the radios are and which sandbox has which |
| 4 | Two-node BER experiments | A first real measurement, end to end, with results you can plot |
Every page follows the same skeleton — what it covers and what you will be able to do afterwards, the devices, images and software it needs, the topology, then step-by-step instructions, verification and cleanup. Each carries a Last verified date, and where a page has not been re-run recently it says so rather than pretending.
Each page also says how you run it, and the words mean specific things:
cosmos-run <description> command does imaging, configuration, the experiment, collection and teardown, including on failure.The best pages are both interactive and automated: see it once by hand, then let the sweep produce the curve.
The testbed's own conventions: the experiment life cycle, imaging nodes with omf load, first contact with an SDR, and how to get results back off a node once you have them. Start here even if you are experienced elsewhere — the reservation and imaging model is specific to this facility.
The largest and most actively maintained group. Software-defined radio from a single tone between two nodes up to full bit-error-rate waterfalls measured on hardware, including an OFDM laboratory at FR1 that compares three different ways of moving an operating point and explains why they disagree. Also spectrum sensing, FPGA-accelerated visualisation with RFNoC Fosphor, and decoding live aircraft transponders. The group index separates the Ansible-orchestrated pages from the older ones and tells you which SDRs live in which sandbox.
Complete cellular stacks you can stand up yourself. Three independent 5G SA implementations — Amarisoft, OCUDU (srsRAN) and Duranta (OAI) — each available both over the air and in a soft-only mode that needs no radio at all, so you can learn the software before booking hardware. Beyond that: O-RAN and ONAP deployment, an end-to-end Split 7.2 with real O-RUs from three vendors, and a tri-stack comparison running all three on one node.
Millimetre-wave work at 28 and 60 GHz. The IBM 28 GHz phased-array modules (PAAM) from first principles through beam steering, MIMO with real-time baseband, and link-rate measurement; Sivers 60 GHz front-ends including 802.11ad preamble processing on RFNoC; and Xilinx RFSoC experimentation. This is where the testbed's most specialised hardware lives.
Measuring the channel itself rather than a link through it. Wideband sounding on X310s and USRP-2974s, and the RENEW sounder for 2×2 TDD MIMO and 16×1 massive MIMO.
The 6–24 GHz mid-band, reached through Pi-Radio up/down-conversion front-ends on sandbox 1. A graded sequence: a sine wave to prove the chain works, analog FM, OFDM BER across five subcarrier modulations — which also runs at FR1, and is the only route to 16-QAM and 64-QAM — and a full Amarisoft 5G NR link over the same hardware. The frequency plan matters more than anything else here, and the sine tutorial explains why before you can get it wrong.
Ansible-orchestrated 802.11: a single access point with a client, and a cell with one AP and many clients. Short, reliable, and a good first look at the orchestration framework because the radio side is familiar.
COSMOS is unusual in having a real optical network alongside the radios. Set up a path on the physical testbed, reproduce the SigComm 2022 path-switching demonstration, or work in the Mininet-Optical emulator with no hardware at all.
The compute side of the testbed: Alveo FPGA acceleration, Intel Movidius inference sticks, ONAP on Kubernetes, Open Source MANO with OpenStack, and container service migration at the mobile edge — which is the one that ties the radio and cloud halves together.
Low-power IEEE 802.15.4 with Contiki OS, on RZRAVEN USB sticks and Crossbow TelosB motes.
Geo2SigMap — building RF signal maps from geographic data with Sionna ray tracing, and comparing them against what the testbed actually measures.
Measuring network performance with iperf — the throughput baseline worth having before you blame the radio.
Written for school students rather than researchers: AM and FM modulation made visible, at roughly grade-8 level, using real radios.
Work that does not fit the categories above, and is kept because it still teaches something: full-duplex wireless with the FlexICoN cancellers, MobilityFirst, outdoor spectrum measurement, and a smart-city edge demonstration with Mask R-CNN.
Something here out of date, or a tutorial you would like to see? The Last verified line at the foot of each page says when it was last actually run.